Reflex-Guard: A Low-Latency Guardrail for LLM Prompt Safety Using Dense Semantic Embeddings

Multi-dimensional Vulnerability Analysis of LM Safety Guards via Role-Goal-Format Decomposition

Current safety guard models for Large Language Models (LLMs) often fail to block complex Jailbreaking attacks because they focus on the structure of a prompt rather than its true intent. To evaluate this problem, we introduce the Role-Goal-Format (RGF) framework, which analyzes how the combination of a user’s persona, their objective, and the text format can bypass safety filters. We compared two small safety guard models: Prompt Guard 2 and Llama Guard 3. The results show that even with the same overall accuracy, they have completely opposite weaknesses. Prompt Guard 2 is easily bypassed by unusual text formats. On the other hand, Llama Guard 3 lacks semantic understanding, missing 76.7% of virus-related attacks. Furthermore, both models exhibit a common vulnerability to contextual breaking; they fail to activate safety alignment when harmful objectives are framed within authoritative and benign personas, such as Teacher, Engineer, and Investigator. These findings highlight that small safety guard models have clear limitations and require multi-layered defenses that understand both text structure and actual intent.

HiRoute: Hierarchical Routed Prompt Tuning for Safety Alignment of Large Language Models

Large language models (LLMs) remain vulnerable to harmful requests and jailbreak attacks. Parameter-efficient safety alignment methods based on prompt tuning typically rely on a single global prompt or externally selected prompt modules. Such static designs struggle to maintain a cross-category safety boundary while generating constructive responses tailored to specific risks and avoiding over-refusal of benign inputs. To address these limitations, we propose HiRoute, an input-adaptive hierarchical prompt-tuning framework that separates category-agnostic safety control from category-specific response guidance. HiRoute first trains a lightweight hierarchical router on representations extracted from a frozen LLM to jointly detect harmful intent and predict multi-label risk scores. It then freezes both the backbone model and the router and uses preference optimization with alternating gradient updates to learn a shared coarse-grained prompt and a set of fine-grained prompt experts as continuous embeddings. At inference time, benign inputs bypass the safety branch, whereas risky inputs are processed using the shared prompt together with a router-weighted mixture of risk-specific prompt experts. Experiments across three instruction-tuned models show that HiRoute achieves high safety rates across multiple safety benchmarks while preserving safe-response helpfulness, reducing over-refusal, and maintaining competitive performance on general-purpose tasks.

BERM: Low-Overhead Prompt-Injection Detection via In-Situ Benign Representation Modeling

Real-Time Detection and Mitigation of Prompt Injection Attacks in LLM-Integrated Enterprise Systems

Large language models (LLMs) embedded in enterprise workflows cannot structurally distinguish legitimate instructions from adversarial ones in the same token stream, making prompt injection OWASP's top LLM risk for two consecutive editions a persistent threat across direct and indirect vectors. This paper presents PromptShield-RT, a layered, real-time, model-agnostic framework combining input normalization and provenance tagging, lexical-heuristic pattern matching, a statistical classifier, structural anomaly features, and calibrated risk fusion, with policy-driven mitigation (allow/sanitize/quarantine/block) and an explainable, adaptive-feedback mechanism for SOC workflows. We construct an original evaluation corpus, SynPI-Bench (n = 450, six categories), and a template-disjoint held-out generalization set (n = 31) with novel phrasings, obfuscation encodings, and adversarial hard-negative benign text. Using template-grouped 5-fold cross-validation, the fused pipeline achieves 92.4% accuracy (F1 = 0.930, AUC = 0.990), outperforming heuristic-only (57.0%) and naive-averaged (59.2%) baselines, while a lexical classifier reaches 85.9% with lower precision. We report a pronounced generalization gap on the held-out set (48.4% accuracy, 90% false-positive rate on hard negatives), quantifying a known limitation of surface-lexical defenses. The pipeline achieves sub-millisecond P95 latency (0.266 ms), within typical 50 ms enterprise SLAs. We situate PromptShield-RT relative to structural, architectural, and guardrail-product defenses, arguing for layered, defense-in-depth architectures, with reproducible code provided.

COMIC: Reference-Aware Safety Gating for Multimodal Large Language Models

Multimodal large language models (MLLMs) are increasingly used to interact with screenshots, scanned documents, diagrams, and other visually grounded inputs. This shift introduces a new safety risk: in many multimodal jailbreaks, neither the prompt nor the image is harmful in isolation. Unsafe behavior emerges only when the model binds an apparently benign operation, such as summarizing, translating, or following, to a localized visual target. This reveals a structural weakness in current multimodal defenses, which largely moderate the prompt-image pair as a whole even though the true security-relevant unit is the grounded operation-target pair produced during dereference. In this work, we identify and analyze this reference-dependent failure mode and show that existing defenses degrade when harmful semantics are localized, activated only after grounding, and dependent on visual reference resolution. To address this problem, we propose COMIC (Context-Operation-Modality-Image-Classifier), a reference-aware pre-generation safety gate for MLLMs. COMIC first infers the requested operation and reference type, constructs candidate targets from OCR and open-vocabulary proposals, grounds plausible referents, and evaluates safety over explicit operation-target pairs. To handle ambiguity conservatively, COMIC combines max-risk aggregation with quality-aware routing before deciding whether to forward or block a request. We evaluate COMIC across multiple open-source MLLMs, localized and broader multimodal jailbreak benchmarks, and benign reference-sensitive settings. The results show that COMIC consistently improves robustness while preserving benign utility and practical efficiency. More broadly, our findings suggest that multimodal safety cannot be enforced reliably without modeling the requested operation, the visual target to which it applies, and the confidence of that grounding.

Safety Targeted Embedding Exploit via Refinement

Safety training for large language models (LLMs) is conducted predominantly in English, leaving uncertain how well safety mechanisms generalize to low-resource languages and mixed-language code-switching. We show that this creates an epistemic gap in which models confidently generate harmful responses for inputs that fall outside the distribution of their safety training. To study this phenomenon, we introduce STEER (Safety Targeted Embedding Exploit via Refinement), a gradient-guided attack that identifies words contributing most strongly to the model's refusal behavior and iteratively translates them into low-resource languages to suppress refusal while preserving harmful intent. Across six open-source 8B-parameter models, STEER achieves attack success rates of up to 93.0% on JailbreakBench and 96.7% on AdvBench, outperforming random code-switching and Greedy Coordinate Gradient (GCG). The resulting prompts also transfer to GPT-4o-mini, achieving a 35.5% attack success rate without requiring access to the target model, suggesting that the underlying weakness is not specific to a single architecture. These findings demonstrate that safety mechanisms aligned primarily on English cannot be assumed to generalize across multilingual inputs. We argue that improving multilingual safety requires broader coverage during alignment and mechanisms that explicitly detect and abstain on out-of-distribution inputs.